Ethernet is winning the war for AI back-end networks, according to new research from Dell’Oro Group.

Just as Broadcom continued its assault on InfiniBand with the launch of its Tomahawk Ultra switch earlier this week, the research firm’s AI Networks for AI Workloads report suggests Ethernet’s battle to dominate the data center-scale fabric space could see it drive nearly $80 billion in data center switch sales over the next five years.

“When we first initiated our coverage of AI Back-end Networks in late 2023, the market was dominated by InfiniBand, holding over 80% share,” said Sameh Boujelbene, VP at Dell’Oro Group. “Despite its dominance, we have consistently predicted that Ethernet would ultimately prevail at scale.”

Where Infiniband once reigned supreme, the rise of Ethernet has been largely driven by industry demand for an open alternative.

The connection technology is effectively only really found in Nvidia networking offerings (as well as Mellanox, prior to Nvidia acquiring the supplier), with its CUDA software stack deeply integrated with InfiniBand, meaning that while vendors have access to a mature ecosystem, they’re locked into using only that.

Ethernet, on the other hand, is an open, industry-standard protocol that the vast majority of engineers from across the industry are well-versed in. The analyst firm suggested that with the hefty costs required to interconnect hundreds of thousands of GPUs, Ethernet wins out over its proprietary rival.

Multiple suppliers have helped it come back fighting, and quickly, with Boujelbene highlighting the speed at which the standard gained ground in the AI back-end networks space.

With Ethernet achieving 800 Gbps speeds and roadmaps extending to 3200 Gbps by 2030, the analyst said Ethernet is “firmly positioned to overtake InfiniBand” across high-performance deployments, matching InfiniBand performance while maintaining its cost and operational advantages.

In addition to speed, growth from GPU-as-a-Service providers like CoreWeave – which was the first vendor to get its hands on Nvidia’s highly anticipated GB300 NVL72 – was cited as a factor helping Ethernet to win out. The provider, along with Lambda and Vultr, were projected by Dell’Oro Group to outpace that of Tier 1 Cloud Service Providers in the next five years.

Boujelbene continued: “Although Ethernet is gaining ground in AI back-end networks, a key question remains: how much of the Ethernet opportunity will be captured by Nvidia versus other switch vendors?

“In 2024, Celestica, Nvidia, and Huawei led the Ethernet segment of the AI Back-end Switch market. However, we anticipate significant share shifts in 2025 as large-scale Ethernet deployments at Meta, Microsoft, Oracle, and other GPU-as-a-Service Providers ramp up – creating opportunities for additional switch vendors such as Accton, Arista, Cisco, Juniper/HPE, Nokia, and others to gain traction.”